arXiv:1910.11988v1 [math.DS] 26 Oct 2019 {A o hr)i admcocycle random a is short) for eete nt rcutbeadfrihdwt h discrete the with furnished and map countable preserving or finite either be iinsnetewr fvnNuano uoaai 90 [ 1950s in automata on Neumann von of work the since dition srnol hnigwt iedeto due a under time individuals with many changing of randomly motions is the describes latter the nthe on tt admdnmclsse.Mr rcsl,let precisely, More system. dynamical random state converg trajectories individual st their initial motion; different of law with noisy systems noninteracting noise-induc of called motions phenomenon counterintuitive a exhibit sys individual an of movement stochastic the dete represents and mer equations different differential distinctly traditional two scienc phenrespectively, are and (RDS) social systems behavior and dynamical dynamical random natural of in representation arise stochastic a that dynamics [ complex 1970s For in developed icantly rmJlnUi.Tewr sprilydn hntefis,sec UW. first, and the when USTC done visiting partially was f is author grant work pa third development The also faculty Univ. Univers was a Jilin author grant, at from CRG fifth PIMS PIMS The the 1257749, parti grant DMS-1444084. and also was NDF PTCS author and PIMS third DMS-1712701 acknowledges The and grant. CRG 11471344, PIMS a by supported cocycles. Linear exponents, punov YCRNZTO NDSRT-IE ICEESAERANDOM DISCRETE-STATE DISCRETE-TIME, IN SYNCHRONIZATION eemnsi yaiswt iceetm tp nadiscr a in steps discrete-time with dynamics Deterministic e od n phrases. and words Key 2010 phenomeno synchronization of study the concerns paper This h rtato a upre yNF rns1313,11431 11371339, grants NSFC by supported was author first The ∗ ( ,ω n, orsodn authors. Corresponding ahmtc ujc Classification. Subject Mathematics σ plctost ilgclntok r lodiscussed. also are networks biological th to partitioning by Applications described is synchronization partial of ycrnzto spoie ntr fteseta subspac spectral the of exponent term in provided is synchronization Abstract. uharno yaia ytmwt nt tt synchronize state finite with system exponent dynamical of random properties a Bo ergodic such multiplicative probabilistic of terms and In random examples. with systems, dynamical random : ) -algebra ω ∈ −∞ 0 Ω θ F a ipemlilct.Frtecs fcutbesaespa state countable of case the For multiplicity. simple has } napoaiiyspace probability a on ecaatrz ycrnzto hnmnni discrete-t in phenomenon synchronization characterize We nadto,frbt ae ffiieadcutbesaespa state countable and finite of cases both for addition, In . .HAG .QA,S WANG S. QIAN, H. HUANG, W. samaual aiyo otnosmpig on mappings continuous of family measurable a is of Ω ycrnzto,Dsrt-ie iceesae Random Discrete-state, Discrete-time, Synchronization, a , iceetm,dsrt-tt admdnmclsystem dynamical random discrete-state discrete-time, 30 aallt h ieo olna yaia ytm theory systems dynamical nonlinear of rise the to parallel ] YAIA SYSTEMS DYNAMICAL A ( n, 1. · : ) rmr 71;Scnay34D06. Secondary 37H15; Primary Introduction extrinsic (Ω S → S , F 1 µ , ∗ , ) .X-.YE X.-F. F. , oss[ noises where , n n n ot uhr eevstn Aadthe and UA visiting were authors forth and ond S ∈ mnsi yaia ytm:Tefor- The systems: dynamical rministic tt e nosnhoie subsets. synchronized into set state e o h nv fAbra n Scholarship a and Alberta, of Univ. the rom = h nue ierccce eshow we cocycle, linear induced the N ooesohsi motion. stochastic one to e lyspotdb SCgat 11771026, grants NSFC by supported ally 31 { orsodn oteLyapunov the to corresponding e over , dsnhoiain h stochastic The synchronization: ed µ s tssnhoieudracommon a under synchronize ates 1 n 1303 te uhr were authors Other 11731003. and 012 .Tednmc fa D may RDS an of dynamics The ]. j tal upre yNECdiscovery NSERC by supported rtially mn.Sohsi rcse and processes Stochastic omena. t fWsigo upre yNSF by supported Washington of ity e with tem } fadol fteLyapunov the if only and if s sapoaiiymauedefined measure probability a is la ewrsa particular as networks olean ∗ oooy o ie measure- given a For topology. omndtriitclwthat law deterministic common ye fmdl htgeneralize, that models of types t-tt pc a ogtra- long a has space ete-state N .YI Y. AND , eoethe denote s ttsia hsc employs physics statistical es, nadsrt-ie discrete- discrete-time, a in n 28 (Ω .Tesbetwssignif- was subject The ]. e hrceiainof characterization ce, , F e,temechanism the ces, m,discrete-state ime, ,θ µ, , intrinsic S yaia ytm,Lya- systems, dynamical tt set state hc aifisthe satisfies which ) .. o each for i.e., , uncertainties; ( hc can which dtds-RDS n . , 2 W.HUANG,H.QIAN,S.WANG,F.X.-F.YE,ANDY.YI cocycle property over the metric dynamical system (Ω, F,µ,θ) (see Section 2.1). A is called a finite-state random dynamical system (finite-state RDS for short) or a countable-state random dynamical system (countable-state RDS for short), when S is a finite set or a countable set respectively. The cocycle A admits a unique matrix representation M, called the induced linear cocycle:
M(n,ω) = (Msisj (n,ω)),
1, si = A(n,ω)sj, (1.1) Msisj (n,ω) = i, j = 1, 2, · · · , (0, otherwise, where n ∈ N, ω ∈ Ω. We will show in Section 2 that M is indeed a linear cocycle over (Ω, F,µ,θ) acting on Rk, if #S = k, and on ℓ1, if #S = ∞ (see Lemmas 2.1, 2.2). In parallel to continuous-state RDS theory, the framework of dtds-RDS is more practical and has enjoyed a wide range of applications in science and engineering [23, 31]. Particularly fitting examples include the random and probabilistic Boolean networks. The random Boolean network, introduced in 1969 by S. Kauffman [16, 17] as a simple model for gene regulatory networks, has two state variables 0, 1 representing “off” and “on” states of a gene respectively. A network of genes evolves according to a given Boolean function. A random Boolean network concerns randomly chosen initial data [5], while a probabilistic Boolean network involves randomly chosen i.i.d. Boolean functions to build a “randomly chosen constituent network” with deterministic Boolean dynamics in a “random period of time” [27]. Since then, different forms of Boolean network dynamics has found applications in neural computations, gene networks, as well as Boltzmann machines that led the current deep learning [1, 8, 12, 21, 32]. To describe a more general random or stochastic Boolean network, we let S be the set of Boolean variables on the nodes of a network, Ω be the probability space assembling all possible randomness, noise, or stochasticity in the network, and α(ω) : S → S, ω ∈ Ω be a measurable family of state transition maps determined by a set of Boolean functions describing the connectivity of the network. Then n A(n,ω)= α(θ −1ω) ◦···◦ α(θ(ω)) ◦ α(ω), n ∈ N defines the random cocycle of the corresponding dtds-RDS and the and the induced linear cocycle M are the adjacency matrices of the network. With such a general setting, not only do we allow any finite or countable number of state variables, but also the randomness involved is also made general which particularly allows the dependency on the past history. We remark that unlike the case of a discrete stochastic process which emphasizes intrinsic stochasticity in the movement of each and every individual, the dtds-RDS modeling approach emphasizes the randomness in the “law of motion" for an entire population of individuals which are governed by the same law [23]. When A is a finite-state RDS, it generates a random subshift of finite type of the S- symbolic random skew-product flow over (Ω, F,µ,θ), with induced cocycle M being the random transition matrices. Indeed, let Σ denote the set SN of all sequences of elements of S endowed with the product topology, together with the left-shift operator T . Then i ΣA(ω)= {(s0,s1, · · · ) : si ∈ S, Msisi+1 (1,θ ω) = 1}, ω ∈ Ω n n become a random subshift of finite type in Σ × Ω, i.e., T ΣA(ω) ⊆ ΣA(θ ω), n ∈ N. We note that when A is not a finite-state RDS, a similar generating symbolic dynamical system is undefined and thus the cocycle A is a more general description of the corresponding dtds-RDS. We note that the above notion of a countable-state RDS allows the dynamical consideration of a wide range of applications including random lattices and infinite networks. A well-developed concept in the random dynamical systems theory is synchronization (see [3, 6, 14, 24] and reference therein), which is intimately related to the “random attractor" in SYNCHRONIZATION IN DTDS-RDS 3 random as well as in non-autonomous dynamical systems. Roughly speaking, synchronization describes the phenomenon that for almost surely two different initial states collapse into a single one after sufficiently long time. This related property also has interests in neuron biol- ogy. In [20], synchronization is well discussed for neuron networks, which are formulated by a system of stochastic differential equations with white noise which is considered as the stimulus to the network. If along almost each single stimulus realization the response of the network always remains the same independent of which initials it starts from (i.e., synchronized), then the neuron network is said to be reliable. Under general conditions, the sign of the maximal Lyapunov exponent is set as a criterion for the reliability of the network: the negativity of maximal Lyapunov exponent implies reliability and positivity implies unreliability [20]. In [31], the notion of synchronization for a dtds-RDS is introduced and investigated under the spirit that an i.i.d dtds-RDS is a more refined model for stochastic systems than its counter part - the Markov chain. Adopting this notion to the present setting, given a dtds-RDS A over (Ω, F,µ,θ) and ω ∈ Ω, a pair {s,s′} ⊂ S is said to ω-synchronize if there exists n(ω) ∈ N such that A(n,ω)s = A(n,ω)s′, n ≥ n(ω). A is said to synchronize if for µ-a.e. ω ∈ Ω, any pair {s,s′} ⊂ S ω-synchronizes. In this paper, refining and generalizing results in [31], we give an equivalent characterization of synchronization for a dtds-RDS from the viewpoint of the multiplicative ergodic theory. Our main results state as follows.
Theorem A. Consider a finite-state RDS A with the k-state set S = {s1, · · · ,sk}. Then the following holds: (i) For any ω ∈ Ω, the Lyapunov exponents of the induced linear cocycle M acting on Rk exist and take at least the value λ1(ω) = 0 and possibly another value λ2(ω)= −∞. (ii) For any ω ∈ Ω, there exists a partition of S :
η(ω)= {W1(ω), · · · ,Wm1(ω)(ω)},
where m1(ω) is the multiplicity of the Lyapunov exponent λ1(ω) = 0, such that a pair ′ ′ {s,s } ⊂ S ω-synchronizes if and only if s,s ∈ Wi(ω) for some integer 1 ≤ i ≤ m1(ω). (iii) The dtds-RDS A synchronizes if and only if for µ-a.e. ω ∈ Ω, M admits precisely two Lyapunov exponents λ1(ω) = 0, λ2(ω) = −∞ with respective multiplicities m1(ω) = 1, m2(ω)= k − 1. For each ω ∈ Ω, the ω-synchronized partition η(ω) stated in Theorem A (ii) can be of course defined through the equivalence relation that si ∼ sj in S if and only if {si,sj} ω-synchronizes. We note that not only does Theorem A (ii) characterize the number of equivalence classes but also the proof of it gives constructive descriptions of the synchronized subsets {Wi(ω)} explaining the mechanism of partial synchronization. In applying Theorem A to a particular finite-state RDS A, of importance is the charac- terization of the multiplicity m1(ω) of the 0 Lyapunov exponent rather than the Lyapunov exponent itself because it is always attained by (i) above. One useful approach in making such a characterization for a particular model with ergodic µ is to combine Theorem A with the multiplicative ergodic theorem (see Theorem 2.1) to show that m1(ω) is a constant for µ-a.e. ω. This then gives a characterization on whether A is synchronized or otherwise its number of synchronized subsets. We will demonstrate such applications in Section 5 with a p53 random network. For a countable-state RDS, the ω-synchronized partition and subsets can be defined through the same equivalence relation, but similar constructive descriptions of synchronized subsets are 4 W.HUANG,H.QIAN,S.WANG,F.X.-F.YE,ANDY.YI not available. In addition, unlike the finite-state case, Lyapunov exponents for a countable- state RDS need not exist in general, and even they do in some special situation, there can be uncountably many of them (see Remark 2.2). Nevertheless, we have the following result.
Theorem B. Consider a countable-state RDS A with the state set S = {si : i ∈ N}. Then the following holds: (i) For any ω ∈ Ω, let η(ω)= {Wi(ω) : i ∈Iω} be the ω-synchronized partition of S, where Iω is finite or countable and for each ′ ′ i ∈ Iω, Wi(ω) is a synchronized subset, i.e., s,s ∈ Wi(ω) if and only if {s,s } ⊂ S ω-synchronizes. Then 1 1 {v ∈ ℓ : λ(ω, v)= −∞} = {v ∈ ℓ : vj = 0, ∀i ∈Iω}, j:sjX∈Wi(ω) where λ(ω, v) is the Lyapunov exponent of the induced linear cocycle M associated with ω, v defined in (2.4). (ii) The dtds-RDS A synchronizes if and only if for µ-a.e. ω ∈ Ω the Lyapunov exponent −∞ is attained and +∞ 1 1 {v ∈ ℓ : λ(ω, v)= −∞} = v ∈ ℓ : vi = 0 . ( i ) X=1 We note that Theorem B (ii) implies that if a countable-state RDS A synchronizes, then the closure of the spectral subspace associated with the Lyapunov exponent −∞ has codimension- 1 for µ-a.e. ω ∈ Ω. The paper is organized as follows. In Section 2, we study the induced linear cycles and their ergodic properties. In particular, Lyapunov exponents are characterized for the finite- state case and Theorem A (i) is proved. Basic notions of random dynamical systems and a general multiplicative ergodic theorem are also recalled. Section 3 is devoted to the analysis of synchronization phenomenon for a finite-state dtds-RDS. We give a characterization of full synchronization with respect to the Lyapunov exponent 0 and a characterization of partial synchronization with respect to synchronized subsets. Theorem A (ii), (iii) are proved. Similar analysis is conducted in Section 4 for the countable-state case. In particular, we give a characterization of full synchronization with respect to the spectral subspace of the Lyapunov exponent −∞ and a characterization of partial synchronization with respect to synchronized subsets. Theorem B (i), (ii) are proved. In Section 5, we give an example of probabilistic Boolean networks, i.e., the p53 random network, to demonstrate applications of Theorem A. Some discussions on similar applications in more complicated random networks are also given.
2. Ergodic properties of dtds-Random dynamical systems In this section, we study basic ergodic properties of the dtds-RDS, and in particular, we show a multiplicative ergodic theorem for the induced linear cocycle. Notions of random dynamical systems, cocycles, and the classical multiplicative ergodic theorem will be recalled for the case of discrete time variable. 2.1. Matrix representations as a cocycle. Let (Ω, F,µ) be a probability space, where µ is a probability measure defined on the σ-algebra F of Ω. (Ω, F,µ,θ) is called a metric dynamical system if θ :Ω → Ω is a measurable, measure-preserving transformation, i.e., µ(θ−nB)= µ(B), n ∈ N, B ∈ F. Let X be a topological space, called state space. A cocycle SYNCHRONIZATION IN DTDS-RDS 5 over (Ω, F,µ,θ) acting on X is a family {Φ(n,ω), n ∈ N, ω ∈ Ω} of continuous mappings on X which are measurable in ω and satisfy the following cocycle property: Φ(0,ω)= id, Φ(m + n,ω)=Φ(n,θmω) ◦ Φ(m,ω), m,n ∈ N, ω ∈ Ω. With the cocycle property, it is clear that the mapping φ : N × X × Y → X × Y : n φ(n,x,ω) = (Φ(n,ω)x,θ ω), n ∈ N, x ∈ X, ω ∈ Ω defines a measurable skew-product flow on the phase space X × Ω, called a random dynamical system. If the state space X is a normed vector space and Φ(n,ω) is a bounded linear operator for each n ∈ N,ω ∈ Ω, then Φ is called a linear cocyle over (Ω, F,µ,θ) acting on X. Now consider the dtds-RDS A described in Section 1. Let M be the matrix representation of A defined in (1.1).
Lemma 2.1. M depends on ω ∈ Ω measurably and satisfies the cocycle property over (Ω, F,µ,θ).
Proof. Denote S = {si} and let Γ = {f : S → S} be the collection of all deterministic self-maps of S. For each f ∈ Γ, define Mf =: ((Mf )ij) as follows
1, si = fsj, (2.1) (Mf )ij = i, j = 1, 2, · · · . (0, otherwise,
It is easy to see that {Mf } satisfies
(2.2) Mf1 · Mf2 = Mf1◦f2 , f1,f2 ∈ Γ. Note that for fixed n ∈ N and ω ∈ Ω, the matrix representation M in (1.1) satisfies M(n,ω)= MA(n,ω). The lemma now follows from (2.2), the measurability of A in ω, and the cocycle property of A. The following lemma shows that M is a linear cocycle acting on a suitable state space. Lemma 2.2. Let A be a dtds-RDS with matrix representation M. (i) If A is a finite-state RDS with k-state set, then M is a linear cocycle over (Ω, F,µ,θ) acting on Rk; (ii) If A is a countable-state RDS, then M is a linear cocycle over (Ω, F,µ,θ) acting on +∞ 1 v +∞ ℓ := { = (vi)i=1 : |vi| < +∞}. i=1 P Proof. (i) Since for each n ∈ N, M(n,ω) is a k × k matrix, M is a linear cocycle acting on Rk by noting that any k × k matrix is a bounded linear operator on Rk with respect to the standard Euclidean norm. (ii) We only need to verify that for any fixed n ∈ N and ω ∈ Ω, MA(n,ω) defined in (1.1) is 1 v +∞ 1 +∞ u v a bounded linear operator on ℓ . For any = (vi)i=1 ∈ ℓ , denote (ui)i=1 = = MA(n,ω) . Then
0, A(n,ω)sj 6= si for all j ≥ 1 u = i vj, otherwise, j:A(n,ω)sj =si P and therefore +∞ ∞ +∞ (2.3) |ui| = kuk1 = | vj|≤ |vi| < +∞. i i i X=1 X=1 j:A(n,ωX)sj =si X=1 6 W.HUANG,H.QIAN,S.WANG,F.X.-F.YE,ANDY.YI
Remark 2.1. We note that for a countable-state RDS A, the induced linear cocycle M needs not define a cocycle acting on the state space ℓp for 1 < p ≤ +∞. For instance, if for some n ∈ N and ω ∈ Ω, it satisfies that A(n,ω)si = s1 for all i = 1, 2, · · · . Then v = (1, 1/2, · · · , 1/n, · · · ) ∈ ℓp for any 1
Quantities λi,mi, i = 1, · · · , r are referred to as the Lyapunov exponents and their multi- plicities, respectively. For cocycles acting on a Banach space, Lyapunov exponents and their multiplicities can be similarly defined, provided that they exist. When restricting to the case of finite-state RDS, we have the following refined result for the Lyapunov exponents which implies Theorem A (i). Proposition 2.1. Let A be a finite-state RDS and M be the induced linear cocycle over (Ω, F,µ,θ) acting on the state space Rk equipped with the standard Euclidean norm k · k, where k =#S. Then for any ω ∈ Ω and any v ∈ Rk, 1 λ(ω, v) := lim log kM(n,ω)vk n→+∞ n exists and equals either 0 or −∞, and consequently, the Lyapunov exponents of M take at most two values λ1 = 0 or λ2 = −∞. Moreover, the Lyapunov exponent λ1 = 0 is always attained. Proof. We note that with n,ω varying, there are only a finite number of choices of M(n,ω). It follows that for any v ∈ Rk, kM(n,ω)vk only take a finite number of different values for all n ∈ N, ω ∈ Ω. Now we let ω ∈ Ω be fixed. If M(n,ω)v6= 0 for all n ∈ N, then kM(n,ω)vk have uniform positive upper and lower bounds, i.e., there exist constants 0 < κ1 < κ2 < +∞ independent of n,ω such that κ1 ≤ 1 kM(n,ω)vk≤ κ2 for all n ∈ N. Thus, lim log kM(n,ω)vk exists and equals 0. n→+∞ n SYNCHRONIZATION IN DTDS-RDS 7
If there exists n ∈ N such that M(n,ω)v = 0, then M(m,ω)v = 0 for any m ≥ n. Thus 1 lim log kM(n,ω)vk = −∞. n→+∞ n We now argue that the Lyapunov exponent λ1 = 0 is always attained. For otherwise, there exists an n ∈ N such that M(n,ω)v = 0 for all v ∈ Rk. This is impossible since M(n,ω) is not a zero matrix.
In the case of countable-state RDS, for any ω ∈ Ω and any v ∈ ℓ1, define 1 (2.4) λ(ω, v) = limsup log kM(n,ω)vk1, n→+∞ n which is referred to as the Lyapunov exponent of M associated with ω, v, whenever the limit exists. Remark 2.2. (i) Unlike Theorem 2.1, the conclusion of Proposition 2.1 holds for all ω ∈ Ω instead of a full µ-measure set. (ii) There have been works in infinite-dimensional multiplicative ergodic theorem [4, 7, 10, 19]. However, all these works are for cases of countably many Lyapunov exponents. In our case, the induced linear cocycle of a countable-state RDS can admit uncountably many Lya- punov exponents. As an example, let A(n,ω)= f n, n ∈ N, ω ∈ Ω where f : S = {si, i ∈ N} → S is a deterministic map such that fsi = si−1 for all i ≥ 2 and fs1 = s1. For any given λ ∈ (0, 1), choose λ v = , −λ, −λ2, · · · , −λn, · · · ∈ ℓ1. 1 − λ Then for the induced linear cocycle M, 1 λ(ω, v) = limsup log kM(n,ω)vk1 n→+∞ n 1 λ λ(1 − λn) λn+1 = lim log − + n→+∞ n 1 − λ 1 − λ 1 − λ 1 2λn+1 = lim log n→+∞ n 1 − λ = log λ.
Thus, for this example, any value in (−∞, 0) is a Lyapunov exponent of M.
3. Synchronization in finite-state RDS In this section, we study synchronization phenomenon for a finite-state RDS A with state set S = {si : i = 1, 2, · · · , k}. Recall that the induced linear cocycle M over (Ω, F,µ,θ) acts on the state space Rk equipped with the standard Euclidean norm. 3.1. A necessary and sufficient condition for synchronization. By Proposition 2.1, for any ω ∈ Ω, M admits at most two Lyapunov exponents: λ1(ω) = 0, λ2(ω)= −∞. Let (3.1) V (ω)= {v ∈ Rk : λ(ω, v)= −∞}. It is easy to see that V (ω) is a linear subspace of Rk. The following result says that V (ω) is actually contained in a co-dimension-1 hyperplane E0. 8 W.HUANG,H.QIAN,S.WANG,F.X.-F.YE,ANDY.YI
Lemma 3.1. For any ω ∈ Ω, k ⊤ (3.2) V (ω) ⊆ E0 =: {v = (v1, · · · , vk) : vi = 0}. i X=1 ⊤ Proof. Let v = (v1, · · · , vk) ∈ V (ω). Then there exists n ∈ N such that M(n,ω)v = 0. If k k ⊤ we denote M(n,ω)v=u= (u1, ..., uk) , then ui = vi. Hence v ∈ E0. i=1 i=1 k P P Note that for any v ∈ R \V (ω), we have λ(ω, v) = 0. Denote m2(ω) = dim V (ω) and k (3.3) m1(ω) = dim R − dim V (ω)= k − m2(ω).
As in Theorem 2.1, we call m1(ω),m2(ω) multiplicities of λ1(ω), λ2(ω) respectively. From Proposition 2.1, we know that
(3.4) m1(ω) ≥ 1, m2(ω) ≤ k − 1. The following result, from which Theorem A (iii) follows, shows that the synchronization of A happens exactly when the equality in (3.4) holds true. Theorem 3.1. A synchronizes if and only if for µ-a.e. ω ∈ Ω, M admits precisely two Lyapunov exponents λ1(ω) = 0, λ2(ω)= −∞ with respective multiplicities
(3.5) m1(ω) = 1, m2(ω)= k − 1. Proof. Suppose A synchronizes. Then for µ-a.e. ω ∈ Ω, there are integers n(ω) ∈ N such that for every m ≥ n(ω), we can find an integer 1 ≤ ℓ(ω,m) ≤ k such that
A(m,ω)si = sℓ(ω,m), i = 1, · · · , k. For given ω and m ≥ n(ω), it follows from (1.1) that the matrix M(m,ω) has every entries on the ℓ(ω,m)-th row being 1 and all other being 0. Let v ∈ E0, where E0 is the co-dimension-1 hyperplane defined in (3.2). We then have M(m,ω)v = 0, i.e., v ∈ V (ω), where V (ω) is the spectral subspace defined in (3.1). Hence E0 ⊆ V (ω), and by Lemma 3.1, we actually have
(3.6) V (ω)= E0.
It follows that m2(ω) = dim V (ω) = dim E0 = k − 1 and λ2 = −∞ is attained. By Proposition 2.1, the Lyapunov exponent λ1 = 0 is always attained with m1(ω)= k −m2(ω)= 1. Now suppose (3.5) holds. Then dim V (ω)= k−1, µ-a.e. ω ∈ Ω. It follows from Lemma 3.1 that (3.6) holds for µ-a.e. ω ∈ Ω. Let si, sj be any two distinct elements of S and denote by ei, k respectively ej, the i-th, respectively the j-th, standard unit vector in R . Since ei − ej ∈ E0, we have by (3.6) that ei − ej ∈ V (ω) for µ-a.e. ω ∈ Ω. It follows from (3.1) that, for a fixed such ω, there exists n(ω) sufficiently large such that for all n ≥ n(ω),
M(n,ω)(ei − ej)= 0, i.e., M(n,ω)ei = M(n,ω)ej, or equivalently, A(n,ω)si = A(n,ω)sj. This show that any pair of elements in S synchronizes, hence A synchronizes. SYNCHRONIZATION IN DTDS-RDS 9
3.2. Synchronization along synchronized subsets. It can be seen from the definition of synchronization that if the cocycle A is invertible, then none pair of elements in S can synchronize. Thus, for a dtds-RDS, total non-synchronization and total synchronization are two extreme situations, and partial synchronization is to be expected in general. Below, we give a characterization of the mechanism of partial synchronization for a finite-state RDS A. We call ξ = {W1, · · · ,Wp} a partition of S if each Wi, referred to as a component of ξ, is p a subset of S, Wi ∩ Wj = ∅ for all i 6= j, and Wi = S. Let PS be the set of all partitions i=1 of S. For ξ, η ∈ PS , we say S ξ 4 η ⇐⇒ each component of ξ is the union of some components of η.
It is clear that the binary relation “4" defines a partial ordering on PS . A partition family Q ⊆ PS is called a chain if for any ξ, η ∈ Q, either ξ 4 η or η 4 ξ. We call η ∈ Q a minimal partition of Q if ξ 4 η and ξ ∈ Q imply that ξ = η. It is a well-known fact that any finite chain admits a unique minimal partition. For the sake of analyzing partial synchronization occurred in the dtds-RDS, we would like to consider partitions of S that are connected to the random cocycle A. For any ω ∈ Ω, it is −1 easy to see that ξn(ω) =: {A (n,ω)si : i = 1, ..., k} is a partition of S for each n ∈ N.
Lemma 3.2. For each ω ∈ Ω, {ξn(ω) : n ∈ N} form a nonincreasing chain of PS , i.e.,
ξ0(ω) < ξ1(ω) < · · · < ξn(ω) < · · · , which is in fact a finite chain.
Proof. For fixed n ∈ N, let W be a component of ξn+1(ω) and denote −1 n Sn = A (1,θ ω)(A(n + 1,ω)W ). It follows from the the cocycle property of A that W = A−1(n,ω) A−1(1,θnω)(A(n + 1,ω)W ) = A−1(n,ω)s.
s∈Sn [ −1 Since W is arbitrary and each A (n,ω)s, s ∈ Sn, is a component of ξn(ω), we have ξn(ω) < ξn+1(ω). Since the chain {ξn(ω) : n ∈ N} is nonincreasing and S is finite, their number of components is a constant for n sufficiently large. Hence {ξn(ω) : n ∈ N} is a finite chain.
Given ω ∈ Ω, let η(ω) be the minimal partition of the chain {ξn(ω) : n ∈ N}, which we refer to as the ω-synchronized partition. Components of η(ω) are called ω-synchronized subsets of A.
Proposition 3.1. Consider the finite-state RDS A. Then for any ω ∈ Ω, a pair {si,sj} in S ω-synchronizes if and only if si,sj lie in a same ω-synchronized subset of A.
Proof. By definition of synchronization, if {si,sj} ω-synchronizes, then A(n,ω)si = A(n,ω)sj for n sufficiently large. It follows that si and sj lie in a same component of ξn(ω) for n sufficiently large. Consequently, si and sj lie in a same ω-synchronized subset of A. The converse is also clear. Remark 3.1. (i) Actually, by Proposition 3.1, we can define the partition η(ω) in a more straightforward way through the equivalence relation that si,sj ∈ S lie in the same component of η(ω) if and only if {si,sj} ω-synchronizes. We derive it through a chain of decreasing partitions as in Lemma 3.2 to give a more intuitive sense of how partial synchronization happens in a finite-state RDS. 10 W.HUANG,H.QIAN,S.WANG,F.X.-F.YE,ANDY.YI
(ii) We note that the concept of ω-synchronization for a dtds-RDS A is defined in a pairwise way. In fact, for the finite-state case, this concept is global, i.e., for a given ω ∈ Ω, if W (ω) is a ω-synchronized subset, then there exists n(ω) ∈ N such that A(n,ω)W (ω) is a single state for all n ≥ n(ω). The following proposition relates the cardinality of the ω-synchronized partition η(ω) and the multiplicity of Lyapunov exponent λ1(ω) = 0. Proposition 3.2. For each ω ∈ Ω,
(3.7) #η(ω)= m1(ω), where m1(ω) is the multiplicity of the Lyapunov exponent λ1(ω) = 0.
Proof. For each ω ∈ Ω, since for all n sufficiently large, the number of components of ξn(ω) is a constant, it must equal to #η(ω). For a such n sufficiently large, we note that #ξn(ω) equals to (k − m2(ω)), where m2(ω) is the multiplicity of the Lyapunov exponent λ2(ω) = −∞. In k fact, by definition of Lyapunov exponents, a vector v ∈ R such that λ2(ω, v) = −∞ if and only if v satisfies a homogeneous system of linear equations with coefficient matrix being M(n,ω). By (1.1), the i-th row of M(n,ω) is non-zero if and only if there exists some j such −1 that A(n,ω)sj = si, i.e., A(n,ω) si is non-empty, which, by the construction of ξn(ω), is a component of ξn(ω). Thus, the number of non-zero rows equals #ξn(ω). Note that for each column of M(n,ω), there is only one entry being non-zero, which means that the number of non-zero rows is exactly the rank of M(n,ω). Let V (ω) be as in (3.1). Then its dimension equals (k − #ξn(ω)), i.e., m2(ω)= k − #η(ω). (3.7) now follows from (3.3). Remark 3.2. We note that Theorem 3.1 is a special case of Propositions 3.1, 3.2 when #η(ω) = 1 for µ-a.e. ω ∈ Ω. Theorem A (ii) follows from Proposition 3.1 and 3.2.
4. synchronization in countable-state RDS In this section, we study the synchronization phenomenon for a countable-state RDS A with state set S = {si : i ∈ N}. Recall that the induced linear cocycle M over (Ω, F,µ,θ) 1 1 acts on the state space ℓ equipped with the ℓ -norm which we denote by k · k1. 4.1. A necessary and sufficient condition for synchronization. Let +∞ 1 (4.1) E0 = v ∈ ℓ : vi = 0 , ( i ) X=1 and F0 = v ∈ E0 : v has at most finitely many components being non-zero . We have the followingn basic facts. o
Lemma 4.1. The following holds. 1 (1) E0 is a closed subspace of ℓ . (2) F0 = E0. 1 Proof. (1) It is obvious that E0 is a subspace of ℓ . To show the closeness, we take any sequence v v (n) +∞ v +∞ { n} in E0 such that n = (vi )i=1 → = (vi)i=1 . Since +∞ +∞ +∞ (n) (n) 0 = lim v = lim v = vi, n→+∞ i n→+∞ i i i i X=1 X=1 X=1 we have v ∈ E0. SYNCHRONIZATION IN DTDS-RDS 11
v +∞ u(n) (n) +∞ (2) For any = (vi)i=1 ∈ E0, let = (ui )i=1 be such that
vi, 1 ≤ i
Theorem 4.1. Consider the countable-state RDS A and ω ∈ Ω. Then any pair {si,sj} ⊂ S ω-synchronizes if and only if 1 (4.2) {v ∈ ℓ : λ(ω, v)= −∞} = E0, where λ(ω, v) is defined in (2.4).
Proof. Suppose (4.2) holds but there exist si0 ,sj0 ∈ S with si0 6= sj0 satisfying
(4.3) A(n,ω)si0 6= A(n,ω)sj0 , ∀n ≥ 0. v +∞ 1 Since ei0 − ej0 ∈ E0 and (4.2) holds, there exists = (vi)i=1 ∈ ℓ such that 1 (4.4) kv − (e − e )k < , i0 j0 1 8 1 (4.5) λ(ω, v) = lim log kM(n,ω)vk1 = −∞. n→+∞ n We note by (4.4) that 7 9 9 7 1 v ∈ ( , ), v ∈ (− , − ), |v | < . i0 8 8 j0 8 8 i 8 i i ,j 6=X0 0 It then follows from (4.3) that +∞ 13 kM(n,ω)vk = | v | ≥ |v | + |v |− |v | > , 1 j i0 j0 i 8 i i i ,j X=1 j:A(n,ωX)sj =si 6=X0 0 which leads to a contradiction to (4.5). Conversely, suppose any pair {si,sj} ⊂ S ω-synchronizes. We first show that 1 (4.6) {v ∈ ℓ : λ(ω, v)= −∞} ⊆ E0. +∞ v +∞ 1 v If not, then there exists = (vi)i=1 ∈ ℓ \E0 such that λ(ω, ) = −∞. Since | vi| =: i=1 a> 0, and any pair {si,sj} synchronizes, there exists m> 0 such that P m ∞ 1 kM(n,ω)vk ≥ | v |− |v | > a, 1 i i+1 2 i i m X=1 =X+1 for all n sufficiently large, which is a contradiction to the fact that λ(ω, v)= −∞. Thus 1 {v ∈ ℓ : λ(ω, v)= −∞} ⊆ E0. (4.6) then follows from Lemma 4.1 (1). Next, we show that 1 (4.7) E0 ⊆ {v ∈ ℓ : λ(ω, v)= −∞}. 12 W.HUANG,H.QIAN,S.WANG,F.X.-F.YE,ANDY.YI
Since any pair (si,sj) ω-synchronizes, for any v ∈ F0, we have
kM(n,ω)vk1 = | vi| = 0, i:Xvi6=0 1 when n sufficiently large. Thus F0 ⊆ {v ∈ ℓ : λ(ω, v)= −∞} and hence 1 F0 ⊆ {v ∈ ℓ : λ(ω, v)= −∞}. Now (4.7) follows from Lemma 4.1 (2). Remark 4.1. For a finite-state RDS A, it follows from the proof of Theorem 3.1 that if A synchronizes, then k E0 = {v ∈ R : λ(ω, v)= −∞}, µ − a.e. ω ∈ Ω, where E0 is the co-dimension-1 hyperplane defined in (3.2). However, for a countable-state RDS that synchronize, a similar identity is no longer true, i.e., it is necessary to take closure in the identity (4.2). To see this, consider the example in Remark 2.2. It is easy to see that the cocycle A in this example synchronizes but for any ω ∈ Ω, {v ∈ ℓ1 : λ(ω, v)= −∞} is not of co-dimension-1 because there are more than two other Lyapunov exponents. Hence it cannot equal to the hyperplane E0 defined in (4.1). 4.2. Synchronization along synchronized subsets. In the case of countable-state RDS A, for each ω ∈ Ω, the ω-synchronized partition η(ω) := {Wi(ω) : i ∈Iω} of S can be defined through the equivalence relation as mentioned in Remak 3.2 (ii), i.e., a pair (si,sj) belongs to a same component of η(ω) if and only if (si,sj) ω-synchronizes. We still call each component of η(ω) as a ω-synchronized subset of A. Differing from the finite-state case, the number of synchronized subsets Iω can be infinite. Also, when restricted to each ω-synchronized subset W (ω), A(n,ω)W (ω) need not be a single state for any finite n. Now by restricting Theorem 4.1 to a same Wi(ω) of η(ω), we have the following result which gives Theorem B (i).
Theorem 4.2. Let A be a countable-state RDS. Then for each ω ∈ Ω, 1 1 {v ∈ ℓ : λ(ω, v)= −∞} = {v ∈ ℓ : vj = 0, ∀i ∈Iω}, j:sjX∈Wi(ω) where {Wi(ω) : i ∈Iω} are the synchronized subsets of A and λ(ω, v) is defined in (2.4).
Remark 4.2. Theorem 4.1 is a special case of Theorem 4.2 when #Iω = 1. 5. Applications to i.i.d random networks In this section, we demonstrate some applications of our theoretical findings to certain biological, i.i.d random networks in describing their synchronization behaviors, in particular in determining their number of synchronized subsets. In fact, as the example below will show, our results can be applied to networks with more general external randomness, e.g., those modeled by Poisson noises if the intensity is very small. Let S denote a discrete state set and Γ collect all maps on S, together with a probability measure Q on Γ. We recall that an i.i.d dtds-RDS is that each time a map from Γ is randomly chosen to act on S according to Q. The metric dynamical system (Ω, F,µ,θ) modeling the noise is simply defined by the probability ∞ Γ space (Ω, F,µ)= 0 (Γ, 2 ,Q) together with the left-shift operator θ. Q SYNCHRONIZATION IN DTDS-RDS 13
2 1 5 SIAH-1 p53 Mdm-2
β-catenin p19/14ARF 3 4
Figure 1. The negative feedback loop of p53. Each gene is denoted as node 1-5. The solid arrows denote stimulatory interactions, whereas the dashed arrows denote inhibitory influences.
We shall consider a particular i.i.d random network - the probabilistic Boolean network model for the biochemical dynamics of regulating protein p53 in biological cells, followed by some discussions in treating i.i.d networks with more complexity. 5.1. The random p53 network. It has been shown that the tumor suppressor, p53, is a crucial protein in multicellular organism that prevents cancer develoment [26]. The working mechanism of p53, proposed by Harris and Levine [13], is described by a negative feedback loop as shown in Figure 1. In response to an external stress signal, the cell cycle enters arrest, apoptosis, cellular senescence, and DNA repair [13, 26]. These events were modeled in [9] by a Boolean network. Without the external stress, p53 is in the low steady state and the network is determined by another set of Boolean functions [11]. We assume an entire population of cells simultaneously experience a same external stress signal that is fluctuating. The dynamics of p53 in different cells then can be modeled by a dtds-RDS. To describe the p53 dynamics using dtds-RDS, we consider the external stress of the cell as the extrinsic noise, which comes from, for instance, the DNA damage due to environmental radiation. It can be properly modeled as a discrete-time Poisson process with small intensity λ as follows. Initially, all cells start from different initial conditions and without external stress, the p53 dynamics of all cells follow a map B. Once the external stress appears after time random T with exponential distribution Exp(λ), p53 dynamics of all cells follow a different map A (Figure 2(A)) and cells will engage in the DNA repair process for a constant period of time Tr, where Tr is much smaller than the expected waiting time 1/λ. Denote the map C as the Tr-th iteration of the map A. Afterwards all cells return back to normal and follow the map B (Figure 2(B)) again until another external stress appears. It is possible that another external stress appears during the repair cycle, but the probability of this happening is usually very small, and even this happens, the cell may enter the cycle of apoptosis. The discrete-time counterpart of Poisson process for small λ is a Bernoulli process, e.g., a Bernoulli shift in the language of dynamical systems. More precisely, the noise probability space is simply ∞ (Ω, F,µ)= ({C,B}, 2{C,B}, {p, 1 − p}), 0 Y where Q := {p, 1 − p} denotes the probability measure on {C,B} with C and B taking the measure p and 1 − p, respectively. Note that p ≈ λ if we discretize the time of the Poisson process by the unit time. Let θ be the left-shift map on Ω. Then the product measure µ = {p, 1 − p}N0 is an ergodic θ-invariant probability measure on (Ω, F) ([29, Theorem 1.12]), and consequently (Ω, F,σ,µ) is a metric dynamical system. In the p53 network as shown in Figure 1, there are five nodes in total and each node only admits two values, 1 or 0, representing the active and inactive state respectively. So the state set S is the binary expansion from 0 to 31 and in total, there are 32 states. In the state 14 W.HUANG,H.QIAN,S.WANG,F.X.-F.YE,ANDY.YI
(A) 1 (B) 1 9 5 9 5 27 17 27 17 19 19 28 25 7 28 25 7 13 13 18 18 24 6 3 24 6 3 12 4 12 14 14 4 26 22 26 22 30 30
0 0 8 21 8 21 16 29 11 20 2 15 16 29 11 20 2 15 10 23 10 23 31 31
Figure 2. (A) The state transition map for the Boolean network correspond- ing to p53 dynamics in response to an external stress. (B) The state transition map corresponding to p53 dynamics in the absence of the external stress. transition maps A and B, we use decimal numbers 0-31 to indicate the gene expression. Note that we do not plot the map C which is just ATr . For each ω ∈ Ω, define A : S → S as follows n A(n,ω)= A(1,θ −1ω) ◦···◦A(1,ω), n ∈ N, where C, if the 0−th position of ω is C; A(1,ω)= (B, if the 0−th position of ω is B. It is easy to see that A satisfies the cocycle property and hence it is an i.i.d dtds-RDS over (Ω, F,µ,θ). Figure 2(A) shows that the map A admits two attractors which are in fact two limit cycles: (5.1) 1 → 5 → 7 → 6 → 22 → 30 → 26 → 24 → 25 → 9 → 1, (5.2) 0 → 20 → 31 → 11 → 0, respectively. Note that the map C has the same attractors as the map A. However, under the map B, the limit cycle (5.1) collapses into a fixed point {6}, which indicates the homeostasis of the cell, whereas the limit cycle (5.2) remains the same. The dtds-RDS A does not synchronize, because two nodes from two different basins of attractions, like 7 and 0, will never synchronize. However, according to Theorem A(ii), A is always partially synchronized, i.e., for any ω ∈ Ω, there exists a partition η(ω) of S such that a pair of states belonging to the same component of η is ω-synchronized, and moreover, the cardinality of η(ω) equals the multiplicity m1(ω) of the Lyapunov exponent λ1(ω) = 0. Using Theorem A(ii), we have the following result.
Proposition 5.1. For µ-a.e. ω ∈ Ω, m1(ω) = 5, i.e., η(ω) consists of 5 ω-synchronized subsets. Proof. Let n Ω∗ = {ω ∈ Ω : the 0-th position of θ ω is B for n = 0, · · · , 14}. By the definition of µ, we have 15 µ(Ω∗) = (1 − p) > 0. It is easy to see that for any ω ∈ Ω, there are at least five components of η(ω), and four of them constitutes the attracting basin of the limits cycle (5.2), which are W1 = {31, 29, 21, 23},W2 = {0, 2, 8, 10},W3 = {11, 15} and W4 = {20, 16}, respectively. If ω ∈ Ω∗, SYNCHRONIZATION IN DTDS-RDS 15 then the map B is applied repeatedly for 14 times to all the 20 states in the basin W5 of the limit cycle (5.1), driving all these states into the state 6. Consequently, any two points in W5 is ω-synchronized. It follows from Theorem A(ii) that
m1(ω) = 5, ω ∈ Ω∗.
Since µ(Ω∗) > 0 and µ is ergodic, it follows from Theorem 2.1 that m1(ω) = 5 for µ-a.e. ω ∈ Ω. We note that, for some other ω, if, as n increases, switchings between maps A and B at the 0-th position of θnω is rather frequent, then it may happen that certain two different states in the limit cycle (5.1) never collapse, i.e., W5 can be further decomposed into different synchronized subsets and m1(ω) > 5. But Proposition 5.1 says that such ω’s are of zero µ-measure. Proposition 5.1 allows one to define, up to a µ-null set of ω’s, the equivalence relation of these gene expressions that si ∼ sj, if and only if they are in the same component of η(ω). To our knowledge, such concepts are new to the community of Boolean networks. The ex- trinsic noise-induced synchronization behaviour described above is in principle different from synchronization in mechanical systems (e.g. coupled-oscillators), because no direct interac- tions between cells in our model is required. It would be interesting to conduct a biological experiment to verify the partial synchronization phenomenon of multiple cells with different gene expressions exposed by the same radiation source. In many realistic models of gene regulations, random perturbations are incorporated by making the network never “get stuck” in any state set. Generally speaking, a random gene perturbation means that any given gene has a small probability of being randomly flipped to a different state, e.g., a gene admitting value 0 can be flipped to value 1, although this only happens with very small probability [27]. Under random gene perturbations, a partially synchronized random network usually becomes a synchronized one since states from different attracting basins may collapse together due to the random flipping effect. In fact, this can be understood from Theorem A that the number of synchronized subsets is equal to the multiplicity of 0 Lyapunov exponent, which, if being bigger than 1, is easily reduced to 1 under certain generic perturbations. 5.2. More general i.i.d networks. In the example of p53 random network, the number of state set is only 32. However, a general i.i.d random network in reality can have tens of thousands state variables. In general, to determine the multiplicity of 0 Lyapunov exponent and the synchronized partition for a complex random network with a huge number of state variables along certain infinite sequence of maps are very difficult tasks. Nevertheless, we show below that the cardinality of the synchronized partitions, i.e. the multiplicity of λ1 = 0 by Theorem A(ii), for a general i.i.d dtds-RDS can be estimated, at least numerically, by using the Markov chain it induces. We recall from [31] that an i.i.d dtds-RDS uniquely induces a Markov chain on S with transition probability P = (pij)1≤i,j≤k, where pij = Q{f ∈ Γ : fsi = sj} and S = {s1, · · · ,sk}. An upper bound of the multiplicity of λ1 can be estimated by the number of recurrent states of the induced Markov chain as follows. Given a transit state, say si, of the Markov chain, almost all sample trajectories of the Markov chain visit si only finite times. Then by our construction of the synchronized subsets of {η(ω)} in section 3.2, along almost every sequence of maps determined by the element ω in Ω, the component of η(ω) corresponding −1 to the pre-image of si, A (n,ω)si, is an empty set for n sufficiently large. Consequently, the number of non-trivial components of each η(ω) is no more than the number of recurrent states of the induced Markov chain. 16 W.HUANG,H.QIAN,S.WANG,F.X.-F.YE,ANDY.YI
As to the lower bound, we will show that the multiplicity of λ1 = 0 is no less than the number of recurrent classes of the induced Markov chain. In fact, algorithms has been devel- oped to figure out the latter ones, e.g., if we treat this induced Markov chain as the digraph G = (V, E), then an efficient algorithm of depth-first search of the digraph could be used to identify the recurrent communicating classes in O(|V |+|E|) time [15]. Recall that a recurrent class of a Markov chain is the set of all recurrent states that can go to each other with positive possibilities. In other words, any two initial states belonging to different recurrent classes will not collapse together along almost every sample trajectories of the Markov chain, i.e, they belong to different synchronized subsets. Then Theorem A(ii) implies that the multiplicity of λ1 is no less than the number of recurrent classes of the induced Markov chain. Inside each recurrent class, however, it is a more delicate issue to determine whether two different states belong to a same synchronized subset, for which, the two-point motion techniques (e.g. [31]) can be used. More precisely, by applying the same sequence of maps determined by an element ω in Ω, we construct two infinite trajectories on S starting from two different initial states. For the i.i.d dtds-RDS, this two-point motion induces a Markov chain on S × S with transition probability W being
W(si,sj)→(sm,sℓ) = Q{f ∈ Γ : fsi = sm,fsj = sℓ}, ∀i,j,m,ℓ ∈ {1, · · · , k}.
Note that when si = sj,sm = sℓ, the transition probability is the same as that of the Markov chain induced by the i.i.d dtds-RDS, and when si = sj,sm 6= sℓ, the transition probability is 0. Therefore, {(s1,s1), (s2,s2),..., (sk,sk)} is a recurrent class the Markov chain induced by W. Furthermore, one can show that the i.i.d dtds-RDS synchronizes if and only if {(s1,s1), (s2,s2),..., (sk,sk)} is the only recurrent class of the Markov chain induced by the two-point motion. If there exists any other recurrent class, then for almost every ω in Ω, the first state of any pair inside the class cannot be in the same component of the synchronized partition η(ω) as the second state. By Theorem A(ii), the multiplicity of λ1 = 0 within such a recurrent class for almost every ω is at least 2. Indeed, we may better estimates the lower bound from this restriction. For instance, if (si,sj), (sj,sℓ) and (si,sℓ) are in the same recurrent class other than the trivial one {(s1,s1), (s2,s2),..., (sk,sk)}, then for almost every ω, si,sj and sℓ should be in three different components of η(ω), i.e., the multiplicity of λ1(ω) = 0 in this case is at least 3. In this way, we give a lower bound for the multiplicity of the 0 Lyapunov exponent. We illustrate the estimation by the following example. This example of 4 states comes from [31]. The deterministic maps to choose in the i.i.d dtds-RDS are
1 → 2 1 → 1 1 → 4 1 → 3 1 → 1 2 → 1 2 → 2 2 → 3 2 → 4 2 → 2 Γ= , , , , . 3 → 4 3 → 3 3 → 2 3 → 1 3 → 3 4 → 3 4 → 4 4 → 1 4 → 2 4 → 3 α1 α2 α3 α4 α5 Note that the maps α , α , α , α are permutations, while α is not. One can assign non- | {z 1} 2| 3 {z4 } | {z } | {z5 } | {z } zero probability mass on each map such that the induced Markov chain is always aperiodic and irreducible. If we consider the two-point motion, then there exists a recurrent communicating class, {(1, 3), (3, 1), (2, 4), (4, 2), (2, 3), (3, 2), (4, 1), (1, 4)}, other than the trivial one. That is, if we start from any pair of these states, for some ω, these two infinite long sequences will never synchronize under this i.i.d dtds-RDS. From our previous arguments, the first state cannot be in the same component as the second one in the partition η(ω), e.g., 1 cannot be in the same component as 3. With this restriction, a possible minimal partition could be SYNCHRONIZATION IN DTDS-RDS 17
η(ω)= {{1, 2}, {3, 4}} , which indicates that the multiplicity of λ1 = 0 for some ω is at least 2. So we give an estimation of the lower bound of the multiplicity. References
[1] D. H. Ackley, G. E. Hinton, and T. J. Sejnowski, A learning algorithm for Boltzmann machine, Cog. Sci., 9, 147–169, 1985. [2] L. Arnold, Random Dynamical Systems, Springer, New York, 1998. [3] P. H. Baxendale, Statistical equilibrium and two-point motion for a stochastic flow of diffeomorphisms, Spatial Stochastic Processes: Festschrift for T. E. Harris (Ed. K. S. Alexander and J. C. Watkins), Birkhäser, Boston, 189–218, 1991. [4] A. Blumenthal, A volume-based approach to the multiplicative ergodic theorem on Banach spaces, Dis- crete Contin. Dyn. Syst. 36(5), 2377–2403, 2016. [5] B. Drossel, Random Boolean Networks, Reviews of Nonlinear Dynamics and Complexity (Ed. H. G, Schuster), 69–110, 2008. [6] F. Franco, B. Gess, and M. Scheutzow, Synchronization by noise, Prob. Th. & Related Fields, 168, 511–556, 2017. [7] G. Froyland, C. González-Tokman, and A. Quas, Hilbert space Lyapunov exponent stability, Trans. Amer. Math. Soc., 2019, to appear. [8] H. Ge, H. Qian, and M. Qian, Synchronized dynamics and nonequilibrium steady states in a stochastic yeast cell-cycle network, Math. Biosci., 211, 132–152, 2008. [9] H. Ge and M. Qian, Boolean network approach to negative feedback loops of the p53 pathways: synchro- nized dynamics and stochastic limit cycles, J. Comp. Biol., 16(1), 119–132, 2009. [10] C. González-Tokman, A. Quas, A concise proof of the multiplicative ergodic theorem on Banach spaces, J. Mod. Dyn. 9, 237–255, 2015. [11] Y. Guo, Z. You and H. Ge, Robustness and relative stability of multiple attractors in a stochastic Boolean network (in Chinese), Sci. Sin. Math., 47, 1831–1852, 2017. [12] J. J. Hopfield, Neural networks and physical systems with emergent collective computational abilities, Proc. Natl. Acad. Sci., 79, 2554–2558, 1982. [13] S. L. Harris and A. J. Levine, The p53 pathway: positive and negative feedback loops, Oncogene, 24, 2899–2908, 2005 [14] Y. L. Jan, Équilibre statistique pour les produits de difféomorphismes aléatoires indépendants, Ann. Inst. Henri Poincare, Sect. B, 23, 111–120, 1987. [15] J. Jarvis and D. Shier, Graph-theoretic analysis of finite Markov chains, in: D. Shier, K. Wallenius (Eds.), Applied Mathematical Modeling: A Multidisciplinary Approach, Chapman & Hall/CRC, 2000. [16] S. A. Kauffman, Metabolic stability and epigenesis in randomly constructed genetic nets, J. Theor. Biol., 22(3), 437–467, 1969. [17] S. A. Kauffman, Homeostasis and differentiation in random genetic control networks, Nature, 224, 177– 178, 1969. [18] Y. Kifer, Ergodic Theory of Random Transformations., Birkhäuser, Basel, 1986. [19] Z. Lian and K. Lu, Lyapunov exponents and invariant manifolds for random dynamical systems in a Banach space, Mem. Amer. Math. Soc. 206, 967, 2010. [20] K. K. Lin, Stimulus-response reliability of biological networks, Nonautonomous and Random Dynamical Systems in Life Sciences, Lecture Notes in Math. Biosci. (Ed. P. Kloeden and C. Pötzsche), 135–161, Springer, New York, 2012. [21] F. Li, T. Long, Y. Lu, Q. Ouyang, and C. Tang, The yeast cell-cycle network is robustly designed, Proc. Natl. Acad. Sci., 101, 4781–4786, 2004. [22] P. Liu and M. Qian, Smooth Ergodic Theory of Random Dynamical Systems, Springer, 2006. [23] Y. Ma, H. Qian, and F. X.-F. Ye, Stochastic dynamics: Models for intrinsic and extrinsic noises and their applications (in Chinese), Sci. Sin. Math., 47, 1693–1702, 2017. [24] J. Newman, Necessary and sufficient conditions for stable synchronization inrandom dynamical systems, Ergod. Theory Dyn. Syst., 38(5), 1857–1875, 2018. [25] V. I. Oseledec, A multiplicative ergodic theorem, Trans. Mosc. Math. Soc., 19(2), 179–210, 1968. [26] J. E. Purvis, K. W. Karhohs, C. Mock, E. Batchelor, A. Loewer and G. Lahav, p53 dynamics control cell fate, Science, 336(6087): 1440–1444, 2012. [27] I. Shmulevich and E. R. Dougherty, Probabilistic Boolean Networks: A Model for Gene Regulatory Net- works, SIAM Press, Philadelphia, PA, 2009. [28] J. von Neumann, The general and logical theory of automata, Cerebral mechanisms in behavior, 1(41), 1–2, 1951. 18 W.HUANG,H.QIAN,S.WANG,F.X.-F.YE,ANDY.YI
[29] P. Walters, An Introduction to Ergodic Theory., Springer Verlag, 1982. [30] S. Wolfram, A New Kind of Science, Wolfram Media, IL, 2002. [31] F. X.-F. Ye, Y. Wang, and H. Qian, Stochastic dynamics: Markov chains and random transformations, Discrete & Cont. Dyn. Sys. B, 21, 2337–2361, 2016. [32] Y. Zhang, M.-P. Qian, Q. Ouyang, M. Deng, F. Li, and C. Tang, Stochastic model of yeast cell-cycle network, Physica D, 219, 35–39, 2006.
Wen Huang: School of Mathematical Sciences, University of Science and Technology of China, Hefei, Anhui 230026, PRC E-mail address: [email protected]
Hong Qian: Department of Applied Mathematics, University of Washington, Seattle, WA 98195, U.S.A E-mail address: [email protected]
Shirou Wang: Department of Mathematical & Statistical Sci, University of Alberta, Ed- monton, Alberta, Canada T6G2G1 E-mail address: [email protected]
Felix X.-F. Ye: Department of Applied Mathematics & Statistics, Johns Hopkins University, Baltimore, MD 21218, U.S.A E-mail address: [email protected]
Yingfei Yi: Department of Mathematical & Statistical Sci, University of Alberta, Ed- monton, Alberta, Canada T6G2G1, and School of Mathematics, Jilin University, Changchun 130012, PRC E-mail address: [email protected]